Scalar propagators on adS space

Harold Erbin∗1
1
Sorbonne Universités, UPMC Univ Paris 06, UMR 7589, LPTHE, F-75005, Paris, France
1
CNRS, UMR 7589, LPTHE, F-75005, Paris, France

January 29, 2014

Abstract
This is intended to be a pedagogical review of scalar field propagator on adS space
in view of adS/CFT correspondence. There no new results but instead it gathers ideas
from the literature and present them in a consistent way, with insights on some difficult
points 1 .

∗ [email protected]
1 This version is still a draft and may contain some mistakes, and some parts need more developments.

1
Contents
1 Equations of motion 4

2 Solutions 5
2.1 Separation of variables and solution in x-direction . . . . . . . . . . . . . . . 5
2.2 Solution for the radial direction (k 2 > 0) and scalings . . . . . . . . . . . . . 6
2.3 Solution for the radial direction (k 2 < 0) and scalings . . . . . . . . . . . . . 8
2.4 Asymptotic behavior and boundary field (k 2 > 0) . . . . . . . . . . . . . . . . 8
2.5 Complete (free) solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

3 Another quantization scheme 10

4 Bulk-to-bulk propagator 12
4.1 General properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
4.2 Explicit form . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

5 Boundary-to-bulk propagator 13
5.1 Link to the bulk-to-bulk propagator . . . . . . . . . . . . . . . . . . . . . . . 14
5.2 General expression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
5.3 Witten’s method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

6 Boundary-to-boundary propagator 17

7 Full solution from Witten’s Green function 18

8 Witten’s diagrams 18

9 Evaluation of the action 18

9.1 Position space: Witten’s Green function . . . . . . . . . . . . . . . . . . . . . 19
9.2 Momentum space: Green function ("Freedman’s" method) . . . . . . . . . . . 20
9.3 Momentum space: Green function (alternative method) . . . . . . . . . . . . 22

10 Cubic interactions 22

11 Comments 23

A Conventions 24
A.1 Basic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
A.2 Wick rotation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24

B Anti de Sitter space 25

B.1 Action and equation of motion . . . . . . . . . . . . . . . . . . . . . . . . . . 25
B.2 Coordinates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
B.2.1 Global . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
B.2.2 Patches . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
B.3 Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
B.4 Boundary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27

2
C AdS/CFT correspondence 28
C.1 The dictionnary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
C.2 Choice of coordinate systems . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
C.3 Correlation functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
C.3.1 1-point . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
C.3.2 2-point . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
C.3.3 3-point . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29

D Special functions 30
D.1 Gamma and beta functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
D.2 Bessel functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
D.3 Modified Bessel functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30

3
1 Equations of motion
In this section we consider a scalar field φ(X) on M = adSd+1 background.
The action for a scalar is
1 √
Z
S=− dd+1 X g (g AB ∂A φ ∂B φ + m2 φ2 ). (1.1)
2
Partial derivatives can be replaced by covariant derivatives Dµ because they are acting on
scalar fields. An integration by part gives
1 √ 1 √
Z Z
S=− dd+1 X g φ(− ∆ +m2 )φ + dd+1 X ∂A ( gg AB φ ∂B φ), (1.2)
2 2

where ∆ is the laplacian on adSd+1 . The second term can be rewritten as a surface integral [7,
app. E]:
√ √
Z Z
Sb = dd+1 x ∂A ( gg AB φ ∂B φ) = dd y γ φ nA ∂A φ, (1.3)
M ∂M

where n is the vector normal to ∂M, γ is the induced metric and y the coordinates on the
A

surface.
The first term in (1.2) gives the Klein–Gordon equation:

(− ∆ +m2 )φ = 0. (1.4)

Using the upper-half plane metric (B.17) one computes the laplacian:
1 √
∆ = √ ∂A ( gg AB ∂B )
g
z d+1
  d+1 2 
Ld+1 z 2 2

L z
= d+1 ∂z ∂z + d+1 2 ∂x
L z d+1 L2 z L
gives

z 2  d−1 
∆= 2
z ∂z (z −d+1 ∂z ) + ∂x2 (1.5a)
L
z2  
= 2 ∂z2 − (d − 1)z −1 ∂z + ∂x2 . (1.5b)
L
If φ satisfies the equation of motion the action (1.2) reduces to the surface term (1.3)
only. Now let decompose the surface in two pieces with whose normal are in the z and x
directions respectively, i.e. split the sum in x and z: the first term will vanish if we assume
that the field vanishes for xµ → ±∞ (as is done usually because we don’t need boundary
data for Minkowski space – we refer to litterature for a proper handling of this), and only
the z boundary contributes:
√
Z z=∞
Sb = dd x γ φ nz ∂z φ

∂M z=ε

where ∂M is just Minkowski. We introduced a cut-off because the induced metric diverges
for z = 0 2 . Note that if the space is bound for some reason (e.g. if there is a black hole in
the interior) then the integration should go from zmin to zmax [32].
2 Note that the precise value of the cut-off z
cut-off is not important since the adS metric is left unchanged
by a rescaling of coordinates, which might bring any value zcut-off the ε [17, p. 17]. We will thus do the
computation with ε and remove it at the end.

4
 Assuming 3
an exponential decay for φ at z → ∞ gives finally
√
Z 
Sb = dd x γε φ nz ∂z φ . (1.6)

∂M z=ε

The induced metric and the normal vector are given in section B:
z L2
nz = , γε = η. (1.7)
L ε2

2 Solutions
The solutions to the Klein–Gordon equation (1.4) are discussed (for example) in [3], [1, sec.
4.4], [32, app. A.1].

2.1 Separation of variables and solution in x-direction

Now we will look for a solution to Klein–Gordon equation (1.4): write the fields φ as

φ(z, x) = f (z)Φ(x) (2.1)

since we can hope to separate variables because of translation invariance in the x direc-
tion [19]. Then (1.4) becomes

z 2  d−1 
− z ∂ z (z −d+1 0
f )Φ) + f ∂ 2
Φ + m2 f Φ = 0
L2
where we noted ∂ 2 = ∂x2 = ∆(d) (laplacian on Minkowski), and by dividing with f Φ one can
separate variables:
z d−1 m2 L2 ∂2Φ
− ∂z (z −d+1 f 0 ) + = = −k 2 (2.2)
f z2 Φ
where k 2 is the norm of a d-dimensional vector k µ . This gives the two equations:

(−∂x2 − k 2 )Φk = 0, (2.3a)

− z d+1 ∂z (z −d+1 ∂z ) + m2 L2 + k 2 z 2 fk = 0, (2.3b)
 

and we have added a k subscript since solutions now depends of this parameter. According
to the discussion of Wick rotation (appendix A.2), the expression in Euclidean space of the
first equation is
(−∂E
2
− k 2 )Φk = 0 (2.4)
The only difference here is that k 2 ≥ 0 which justify the sign of the right hand side in (2.2).
Notice also that this equation is different from the one we got in the appendix (opposite
"mass"), which allow plane wave-like solutions: in euclidean space, plane waves are possible
for at most d − 1 directions.
Since one get modes depending on a parameter k, the full solution will be obtained by
superposing all of them 4 : Z
φ(z, x) = dd k fk (z)Φk (x). (2.5)

Before solving explicitly the equations, let’s summary what are the consequences of the
k 2 sign, since solutions will depend on it [34, p. 16], [19, lec. 15], [29]:
3 Thiswill be shown in next section.
4 One could decide to write an overall coefficient φ0 (k) for each mode, as it is done in [32, app. A], but
we prefer to include it in fk .

5
 • k 2 = µ2 > 0 (Euclidean): this will lead to the Euclidean Green function. In the z
direction this gives real exponentials.
• k 2 = −µ2 < 0 (timelike Minkowskian): the momentum satisfies the usual on-shell mass
condition and so µ would be the mass in x-space. The z-equation leads to imaginary
exponentials, corresponding to retarded/advanced Green functions.
• k 2 = µ2 > 0 (spacelike Minkowskian): here the momentum is off-shell. The z solution
are again real exponentials.
If we wanted to interpret µ2 as the mass of a particle in d-dimensional space in the first two
cases 5 , then the second equation (2.3b) should give discrete value for k 2 (i.e. a dispersion
relation). Sadly we will see that it is not the case, so the parameter µ can not be interpreted
as a mass [9, sec. 2.1.1] 6 .
We will work mostly in euclidean space but there is no great difference with Minkowski,
as explained for example in [27].
In (2.3a) (or its Euclidean version) one recognizes the Klein–Gordon equation in d-
dimensional spacetime, whose solutions are plane-waves:
eikx
Φ(x) = . (2.6)
(2π)d
So let’s superpose modes using equation (2.5):

dd k
Z
φ(z, x) = fk (z) eikx (2.7)
(2π)d
and we get that φ(z, x) is the Fourier transform of fk , which is understable due to the
translation invariance in x direction. By inverting the transformation we see that fk (z) is
the solution in momentum space.

2.2 Solution for the radial direction (k 2 > 0) and scalings

Results are identical for spacelike Minkowskian and Euclidean cases, writing k for both [32,
app. A.2.1].
The z equation can be written

z 2 fk00 − (d − 1)z fk0 − (m2 L2 + k 2 z 2 )fk = 0, (2.8)

and it is almost the (modified) Bessel equation (D.6): let’s do the change of variable fk =
z d/2 gk to get 7 [1, sec. 4.4]
 2 
d
z 2 gk00 + zgk0 − + k 2 z 2 + m2 L2 gk = 0. (2.9)
4
Then one does not takes gk as a function of z, but as a function of kz to get (due to
derivatives) 8  2 
d
(kz)2 gk00 + (kz)gk0 − + m2 L2 + k 2 z 2 gk = 0. (2.10)
4
5 Indeed if k 2 = −µ2 < 0, then the previous condition says that Φ satisfies (− ∆ +µ2 )Φ = 0 in

Minkowskian space.
6 If z is bound by above in the interior, then conformal symmetry is broken and µ will take discrete

values [27].
7 To find the power of z, write z α and choose α in order to have z as coefficient for g 0 .
k
8 Here we note k = |k| = µ, which should not be confused with the vector k µ itself; this should be clear

from the context.

6
And from the appendix D.3 on modified Bessel functions, one reads the solution for gk :
gk (kz) = ak Kν (kz) + bk Iν (kz), (2.11)
where we have defined the parameter ν as
r
d2
ν= + m2 L2 . (2.12)
4
Finally the solution for fk is
fk (z) = ak (kz)d/2 Kν (kz) + bk (kz)d/2 Iν (kz). (2.13)
We have to impose that solutions are regular everywhere in the interior, and more specif-
ically for z → ∞. Using asymptotic form (D.7b) for the Bessel functions
Iν (z) ∼ ekz , Kν (z) ∼ e−kz , (2.14)
one sees that Iν diverges so bk = 0 and
fk (z) = ak (kz)d/2 Kν (kz). (2.15)
Using the asymptotic forms of the Bessel functions (D.10), one finds that near the bound-
ary z ≈ 0 the solution behaves like
 ν  ν 
d/2 Γ(ν) 2 Γ(−ν) kz

fk (z) ≈ ak (kz) +
2 kz 2 2
and after simplication
fk (z) ≈ φ0 (k)z ∆− + φ1 (k)z ∆+ (2.16)
where we have defined
φ0 (k) = ak 2ν−1 Γ(ν) k ∆− , φ1 (k) = ak 2−(ν+1) Γ(−ν) k ∆+ . (2.17)
The scaling exponents ∆± are defined as
r
d d d2
∆± = ± ν = ± + m2 L2 . (2.18)
2 2 4
Note that the positivity of the square-root gives the Breitenlohner–Freedman bound [34,
p. 16]
d2
m2 L2 > − . (2.19)
4
For future reference note that we have
 2ν
φ1 (k) Γ(−ν) k φ1 (−k)
= = (2.20)
φ0 (k) Γ(ν) 2 φ0 (−k)
In position space (2.7) the asymptotic (2.16) becomes
φ(z, x) ≈ φ0 (x)z ∆− + φ1 (x)z ∆+ (2.21)
where φ0 (x) and φ1 (x) are the Fourier transform of φ0 (k) and φ1 (k).
Another way to deduce these scalings are to plug the ansatz f (z) = z ∆ into (2.3b) [27]:
− ∆(∆ − d) + m2 L2 + k 2 z 2 z ∆ = 0. (2.22)

Close to the boundary z ≈ 0 and one can ignore the term in z 2 which otherwise spoils the
power-law solution, and then one finds again the two roots (2.18).

7
2.3 Solution for the radial direction (k 2 < 0) and scalings
For a timelike Minkowskian momentum all the previous results are found through the re-
placement µ → iµ [34, p. 16]. So the differential equation becomes the one for Bessel
functions (cf appendix D.2):
 2 
d
(kz) gk + (kz)gk −
2 00 0
+ m L − k z gk = 0.
2 2 2 2
(2.23)
4

Note that here k = iµ (doing the replacement does not change any sign in the first two
terms).
Solutions are J±ν (kz) ∼ K±ν (ikz) if ν is not an integer, Jν (kz) and Yν (kz) other-
wise [3] 9 . Using the previous solution one sees that

K±ν (ikz) ∼ e±ikz (2.24)

so both solutions are regular at z = ∞ and we have to impose in-falling or outgoing boundary
solution, which are related to retarded/advanced Green functions [29], [32, sec. 3.2].

2.4 Asymptotic behavior and boundary field (k 2 > 0)

Now study in more details the asymptotic form (2.16) of the solution:

fk (z) ≈ φ0 (k)z ∆− + φ1 (k)z ∆+ . (2.25)

Since ∆+ > 0 the solution z ∆+ −−−→ 0 is a normalizable solution and corresponds to a

z→0
bulk excitation which decays at the boundary [16, sec. 5.1] and [14, p. 6]. But the other
solution does not decay since ∆− < 0 and it is said to be non-normalizable: it defines a field
on the boundary [8, p. 54] 10
φ0 (k) = lim z −∆− fk (z), (2.26)
z→0

or written in position space through (2.7)

φ0 (x) = lim z −∆− φ(z, x) (2.27)

z→0

where
dd k
Z
φ0 (x) = φ0 (k) eikx . (2.28)
(2π)d
Note that usually one defines the boundary field through (2.27) by using the trick at the
end of the subsection 2.2, just saying that since the divergences is only a power-law in z,
then there should be a x-field. Then later one identifies this boundary field with the one we
got in our development (2.16) 11 .
In principle one asks only for normalizable modes in order to construct the Hilbert
space of the theory, since they corresponds to the physical modes which propagates in the
bulk [3, 4], [10, sec. 6]. On the other side, non-normalizable modes are necessary to specify
boundary conditions, and then these modes do not fluctuate and they provides the classical
9 Hankel functions – the linear combinations J ± iY – can be more suitable for some purposes, since
ν ν
they describe in- and outgoing waves at z = ∞ [29, 32].
10 Due to the fact that m2 can be negative there is some subtleties with the normalizable states which

are not reviewed here.

11 Moreover sometimes one uses the solution in terms of Bessel function (2.13) but before asking to

regularity and this leads to errors in the analysis since one uses only a part of the asymptotic of Bessel
functions, namely Kν ∼ z ∆− and Iν ∼ z ∆+ .

8
background on which the normalizable modes propagate. Thus we should not throw away
the non-normalizable modes.
In the first case it is then necessary to obtain the full solution by using the method of
Green function, since the source gives a response.
Using a cut-off to remove the limit process, this relation can be inverted to give

φ(ε, x) = ε∆− φ0 (x), fk (ε) = ε∆− φ0 (k). (2.29)

This is simply what we have already written in the asymptotic expansion of fk . This last
relation is useful when one wants to compute on the boundary in order to avoid divergences.
It is then possible to verify directly [25, p. 7][28, sec. 3.2] that φ0 is an operator with
conformal dimension ∆− :

φ0 (λx) = lim z −∆− φ(z, λx) = λ−∆− lim (λ−1 z)−∆− φ(z, λx)
z→0 z→0

=λ −∆−
lim (z ) 0 −∆−
φ(λz , λx) = λ−∆− lim
0
(z 0 )−∆− φ(z 0 , x)
z 0 →0 0 z →0

i.e.
φ0 (x) = λ∆− φ0 (λx). (2.30)
But now that there is on the boundary a field which acts as a source for the field in the
bulk, we have to propagates it using Green function (more specifically the bulk-to-boundary
propagator). This object will be computed in a later section and at this point we will come
back to the full solution. This will be quivalent to the full solution we derived in next section
(in fact we will identify the Green function using our solution, replacing the boundary field
by a point-like source).

2.5 Complete (free) solution

Plugging the explicit solution (2.15) for fk into the Fourier transform (2.7) gives the full
solution in position space:

dd k
Z
φ(z, x) = ak (kz)d/2 Kν (kz) eikx . (2.31)
(2π)d

The value of ak is determined by boundary conditions at z = ε (the cut-off is necessary

to avoid divergences) [1, sec. 4.4] using the formula (2.26) 12

φ0 (k) = ε−∆− fk (ε) = ε−∆− ak (kε)d/2 Kν (kε). (2.32)

which gives after inversion

ε∆− −d/2
ak = φ0 (k). (2.33)
k d/2 Kν (kε)
Then fk is given by
 z d/2 K (kz)
ν
fk (z) = ε∆− φ0 (k). (2.34)
ε Kν (kε)
Replacing φ0 (k) by (2.29) gives
 z d/2 K (kz)
ν
fk (z) = fk (ε) (2.35)
ε Kν (kε)
12 Defining the boundary condition at z = ε instead of z = 0 is more logical because it will avoid many

ε factors in latter expression, but it is also necessary to get correct formula when evaluating the actions [4,
app. A].

9
and one sees that taking z = ε gives the right answer.
An approximation of ak can also be found using (2.17):

φ0 (k)
ak ≈ . (2.36)
2ν−1 Γ(ν)k ∆−

Using (2.16) this also leads to the approximate form

φ0 (k)z ∆− + φ1 (k)z ∆+
fk (z) ≈ . (2.37)
φ0 (k)ε∆− + φ1 (k)ε∆+
The final solution in position space becomes
 z d/2 Z dd k Kν (kz) 0
φ(z, x) = ε∆− dd x0 φ0 (x0 ) eik(x−x ) (2.38)
ε (2π) Kν (kε)
d

where we recall that φ0 (x) is the Fourier transform of φ0 (k) (2.28):

dd k
Z
φ0 (x) = φ0 (k) eikx . (2.39)
(2π)d

3 Another quantization scheme

There is a loophole in the previous discussion, where we considered only the case where z ∆+
is normalizable. In the mass range

d2 d2
− < m2 < − + 1 (3.1)
4 4
the action is also finite for z ∆− (with adapted boundary conditions) [18, 3, 22], such that it
exists two different quantizations 13 .
We reproduce the analysis from [18]: consider a field whose asymptotics is

φ ∼ z ∆ φ1 (x) (3.2)

where ∆ is a solution of the equation

∆(d − ∆) + m2 = 0, (3.3)

i.e. r
d d2
∆± = ± + m2 . (3.4)
2 4
We inject this form into the action (1.1)
1 √
Z
S1 = − dd+1 X g (g φ1 B ∂A φ ∂B φ + m2 φ2 ) (3.5)
2
whose integrand is
√
L1 = g g zz (∂z φ)2 + φ02 m2 φ2

m2 m2 2
   
= −z −d+1 (∂z φ)2 + 2 φ2 + z 2∆ φ021 = −z −d+1
(∂z φ)2
+ φ + z 2∆ 02
φ 1
z z2
= −z −d+1 ∆2 z 2∆−2 + m2 z 2∆−2 φ21 + z 2∆ φ02
1 = −z
2∆−d−1
(∆2 + m2 )φ21 + z 2∆ φ02


1 ,
13 Historically this was shown from energetic considerations, see e.g. [6].

10
(where φ0 ∂µ φ and the square hides a g µν ) and using the definition of ∆, the integrand is
L1 = φ21 ∆(2∆ − d)z 2∆−d−1 + z 2∆ φ02
1 . (3.6)
The last term is subleading and the integral will converge if
d
2∆ − d − 1 > −1 ⇐⇒ ∆ > . (3.7)
2
In this range we have 2∆ − d > 0 and the coefficient of L can never vanish. We conclude
that only ∆+ is valid.
But we can also use the integrated form (1.2)
1 √
Z
S1 = − dd+1 X g φ(− ∆ +m2 )φ, (3.8)
2
then throwing the boundary term and using the expression for the laplacian (1.5) we have
the integrand
d−1 m2
 
L2 = z −d+1
φ −∂z +
2
∂z − ∂x + 2 φ
2
z z
2 d − 1 ∆−1
 
=z ∆−d+1
−∆(∆ − 1)φ1 z
2 ∆−2
+ ∆φ1 z + m φ1 z
2 2 ∆−2 00 ∆
− φ1 z
z
= z ∆−d+1 φ21 − ∆(∆ − 1) + ∆(d − 1) + m2 z ∆−2 − φ001 z ∆ .
 

The coefficient in front of z ∆−2 vanishes because of the definition of ∆, and thus
L2 = φ001 z 2∆−d+1 . (3.9)
The integral is convergent if
d
2∆ − d + 1 > −1 ⇐⇒ ∆ > − 1. (3.10)
2
The divergent piece comes from the boundary as can be verified. In this context both ∆±
are valid 14 .
In order to translate these bounds on the mass, we write
r  
d2 d
− + m 2 = ∆− − > −1
4 2
r
d2
+ m2 < 1
4
and thus
d2
m2 < − + 1 ⇐⇒ ν < 1. (3.11)
4
Above this limit we can not use the boundary condition with ∆− .
In the rest of the text we focus on ∆+ boundary condition. We found that the asymptotic
field can be written
φ = z ∆− φ 0 + z ∆+ φ 1 (3.12)
and that φ0 is interpreted as a source on the boundary. But it is not necessary to do again
all the computations: as explained in [18, p. 11–12][22], φ0 and φ1 are related by a canonical
transformation and we can obtain the needed quantities by a Legendre transformation 15 .
The link between boundary terms, boundary conditions and the two quantization schemes
is further explored in [21].
14 In view of the adS/CFT correspondence the bound ∆ > d/2 is strange because we know that the

unitary bound for a scalar field in a CFT is ∆ > d/2 − 1, exactly what is found here.
15 Using an hamiltonian analysis we can show that φ is the canonical momentum associated to φ .
1 0

11
4 Bulk-to-bulk propagator
4.1 General properties
Now let’s look for propagators. The bulk-to-bulk propagator G(X; X 0 ) is defined by
i
(− ∆ +m2 )G = − √ δ (d+1) (X − X 0 ) (4.1)
g

and so it is i times the Green function. The delta functions of the right hand side is defined
such that Z
dd+1 X δ (d+1) (X − X 0 ) φ(X 0 ) = φ(X) (4.2)

(no metric determinant). And then Klein–Gordon equation with source J

(− ∆x +m2 )φ = J (4.3)

is solved by the convolution of G with the source

Z
φ(X) = dd+1 X 0 G(X; X 0 )J(X 0 ), (4.4)

since
√
Z
(− ∆x +m2 )φ = (− ∆X +m2 ) dd+1 X 0 g G(X; X 0 )J(X 0 )
Z
= dd+1 X 0 δ (d+1) (X − X 0 ) G(X; X 0 )J(X 0 ) = J(X).

By doing a Wick rotation we get (where X 0 is now the euclidean time) [23]
1
(− ∆ +m2 )G = √ δ (d+1) (X − X 0 ) (4.5)
g

because the delta function changes by −i and the metric determinant by −1. Beginning
from now we will use only euclidean time (and without writting differently quantities) 16 .
Now for any solution φ(X) of the homogeneous equation (1.4) we have the relationship [5,
p. 67]
√ 
Z
φ(X) = dd+1 X 0 g φ(X 0 )(− ∆X 0 +m2 )G(X; X 0 )
 (4.6)
− (− ∆X 0 +m2 )φ(X 0 )G(X; X 0 ) .

Integrating by part gives (variables are omitted)

√
Z
φ(X) = − dd+1 X 0 g ∂ A (φ∂A G − G∂A φ) (4.7a)
ZM
√
=− dd y 0 γ (φ nA ∂A G − G nA ∂A φ). (4.7b)
∂M

where again γ is the induced metric and y the coordinates on ∂M. From this last form we
can deduce that:
• if G vanishes on ∂M then φ(x) is given by Dirichlet conditions on φ(y);
16 The computation for Minkowski space are mostly the same: there are some modifications due to the

normalizable modes present in Minkowski space [4].

12
 • if nA ∂A G vanishes on ∂M then φ(x) is given by von Neumann conditions on nA ∂A φ(y);
• if none vanishes one has mixed boundary conditions.
Our interested will be in the first case since we found that for our solution φ approaches a
constant φ0 on the boundary (after a rescaling). If φ is given by Dirichlet data then the
solution is unique if µ2 > 0 [5] – see also [33].

4.2 Explicit form

Since we know the modes of the scalar fields in (2.13)

fk (z) = (kz)d/2 Kν (kz), Iν (kz) , (4.8)



we can use them to construct the Green function [23, 24]

dd k
Z 
0 d/2 −ip(x−x0 )
G0 (X, X 0 ) = (zz ) e θ(z − z 0 )Kν (kz)Iν (kz 0 )
(2π)d (4.9)

+ θ(z 0 − z)Iν (kz)Kν (kz 0 ) .

Although one should note that the boundary condition is at z = 0 instead of z = ε.

Question 4.1 Why it is this specific order and why we have the non-normalizable
modes?

This last expression can be integrated to give the result in term of the hypergeometric
function [8, sec. 6.3]
  ∆+ 
2C∆+ ∆ ∆ 1

ξ
G(X; X ) =
0
F , + ; ν + 1; ξ 2
(4.10)
ν 2 2 2 2
where
2zz 0
ξ= (4.11)
z2 + z 02 + (x − x0 )2
(C∆+ will be given later, see (5.18)).
Now we want to fix the boundary at z = ε; we get

dd k Iν (kε)
Z
0
Gε (X, X ) = G0 (X, X ) +
0 0
(zz 0 )d/2 e−ip(x−x ) Kν (kz)Kν (kz 0 ) . (4.12)
(2π) d Kν (kε)

5 Boundary-to-bulk propagator
We will be interested in the boundary-to-bulk propagator K(z, x; x0 ) which happens when
√
a point source is located on the boundary. Due to the fact that γ ∝ z, the right hand side
of the Green equation vanishes, giving

(− ∆ +m2 )K(z, x; x0 ) = 0. (5.1)

The propagator should become a delta function when positions coincide and this will be
proven later.

13
 Here again general solution is given by convolution with a source φ0 (x):
Z
φ(z, x) = dd x0 K(z, x; x0 )φ0 (x0 ). (5.2)

This convolution in position space becomes simply a product in momentum space:

fk (z) = Kk (z)φ0 (k). (5.3)

5.1 Link to the bulk-to-bulk propagator

The formula (4.7) is helpful to get an expression depending of the bulk-to-bulk propagator,
because if we take ∂M as the boundary of the adS, i.e. z = 0, then the φ(y) in the first term
is located on the boundary, and so φ(z, x) is given by the convolution with the propagator.
Using (5.2) and changing the notations to y → x0 , φ(y) → z ∆− φ0 (x0 ) let us identify the
boundary propagator to be
√ 0
K(z, x; x0 ) = lim
0
γz0 (z 0 )∆− nz ∂z0 G(z, x; z 0 , x0 ) (5.4a)
z →0

=L d−1
lim (z 0 )−∆+ z 0 ∂z0 G(z, x; z 0 , x0 ), (5.4b)
z 0 →0

using the normal vector (B.19), and the propagator vanishes for z → ∞.
Now we wants to relate both propagators without derivatives using Green’s theorem [19]
√ 
Z 
dd+1 X g φ(− ∆ +m2 )ψ − (− ∆ +m2 )ψ φ
M
(5.5)
d √
Z
=− d y γ (φ n · ∂ψ − ψn · ∂φ)
∂M

with φ(X) = G(X; X ), ψ(X) = K(X; X 00 ); with G normalizable:

0

z∂z G(X; X 0 ) = ∆+ G(X; X 0 ) (5.6)

(in the following it is understood that only the first argument of K has a z-dependence).
The left hand side gives
√ 
Z 
dd+1 X 00 g G (− ∆ +m2 )K −(− ∆X 00 +m2 )G(X 00 ; X) K(X 00 ; X 0 )
| {z }
=0
√
Z
=− dd+1 X 00 g δ (d+1) (X − X 00 )K(X 00 ; X 0 ) = K(X; X 0 )
M

and the right hand side is:

√
Z  
00
− dd x00 γ nz G(X 00 ; X 0 )∂z00 K(X 00 ; X 0 ) − K(X 00 ; X 0 )∂z00 G(X 00 ; X 0 ) 

z 00 =0
Z  
= − Ld−1 dd x00 (z 00 )−d G(X 00 ; X 0 )z 00 ∂z00 K(X 00 ; X 0 ) − K(X 00 ; X 0 )z 00 ∂z00 G(X 00 ; X 0 )  00

z =0
Z  
= − Ld−1 dd x00 (z 00 )−d ∆− G(X 00 ; X 0 )K(X 00 ; X 0 ) − ∆+ K(X 00 ; X 0 )G(X 00 ; X 0 ) 

z 00 =0
Z   
= − Ld−1 dd x00 (z 00 )−d ∆− − ∆+ G(X 00 ; X 0 )(z 00 )∆− δ (d) (x00 − x0 ) 00 ,

z =0

using (5.20). Equalling both sides gives finally

2ν
K(z, x; x0 ) = lim G(z, x; z 0 , x0 ). (5.7)
z 0∆+
0 z →0

It is evident that generally one will use a cut-off ε in the limit.

14
 Question 5.1 Why G is normalizable? Cf for example [28].

The last formula allows us to find directly the expression [27]

  ∆+
z
K(z, x; x ) = C∆+
0
(5.8)
z 2 + (x − x0 )2

from (4.10) using the limit of the hypergeometric function when z 0 goes to zero [30, p. 24]

∆ ∆ 1
 
F , + ; ν + 1; 0 = 1. (5.9)
2 2 2

5.2 General expression

We can interpret the solution for φ(z, x) (2.38) as given by a point source located at the
position x0 on the boundary, so we can infer the propagator [1, sec. 4.4]
 d/2 Z dd k K (kz)
∆− z ν 0
K(z, x; x ) = ε
0
eik(x−x ) . (5.10)
ε (2π) Kν (kε)
d

In the same way as the field, we have to rescale it as we approach the boundary, where the
propagator becomes a delta function:

dd k ik(x−x0 )
Z
ε−∆− K(ε, x; x0 ) = e = δ (d) (x − x0 ).
(2π)d

To get the propagator in momentum space one has just to remove the field φ0 (k) from
the solution (2.34) in k-space (φ0 (k) = 1 for point source):
 z d/2 K (kz)
ν
Kk (z) = ε∆− . (5.11)
ε Kν (kε)
We note that for z = ε it is equal to one, corresponding to a delta function.
It is possible to find the boundary behavior of Kk using the equation (5.3) and the
asymptotic form (2.16) of fk (z):

fk (z) φ1 (k) ∆+
Kk (z) = = z ∆− + z . (5.12)
φO (k) φ0 (k)

5.3 Witten’s method

Another simpler expression can be discovered thanks to a Witten’s trick 17 [33, sec. 2.4 and
2.5] and [19, lec. 14]: let’s pick the point at z = ∞. It is possible if we consider the space as
compactified, which means that it is just another point (as compactifying Rd to S d ): then
the x part of the metric simplify since

L2
ηµν −−−→ 0, (5.13)
z2 z→∞
17 We can summarized this method: 1. solve Laplace equation, 2. show it is singular at some points, 3.

singular means sources, 4. check what kind of sources it is and if they are physical. In the case of Witten’s
solution, we will see it is a point source.

15
and the space shrinks to a point. For this reason the Green equation (5.1) becomes

− z d+1 ∂z (z −d+1 ∂z ) + m2 L2 K(z) = 0. (5.14)

 

We recognizes the equation (2.3b) for the free scalar field but without the term k 2 z 2 . Then
we have seen that solutions are power-law.

Question 5.2 Do we take a neighbourhood of z = ∞ and not strictly z = ∞? Because

it is weird to have derivatives if we take the function to be valued at one point.

We do the same ansatz as the one for fk and find that

r
d d2
K(z) = C∆+ z , ∆+ = +
∆+
+ m2 L2 , (5.15)
2 4
keeping only the largest root ∆+ from (2.18).

Question 5.3 Why are we keeping only the largest root? Further we find that
z −∆− K → δ, which is the same condition as for fields.

Now we can do an inversion followed by a x-translation (isometries of adS spacetime)

z
z −→ , x −→ x − x0 , (5.16)
z2 + x2
where x2 = ηµν xµ xν , to find the boundary propagator
  ∆+
z
K(z, x; x0 ) = C∆+ . (5.17)
z 2 + (x − x0 )2

Now we want to prove that K is singular only at the point x = x0 on the boundary 18 [1,
sec. 4.4], and thus approaches a delta function, so we computes the integral over x (and
setting x0 = 0 due to translation invariance):
1
Z Z
dd x K(z, x) = C∆+ z ∆+ dd x 2
(z + x2 )∆+
Z ∞
rd−1
= C∆+ z ∆+ Ωd−1 dr 2
0 (z + r2 )∆+
Z ∞
td−1
= C∆+ z d−∆+ Ωd−1 dt
0 (1 + t2 )∆+

with the change of variables t = r/z,

∞ d
1 x 2 −1
Z
= C∆ z d−∆+ Ωd−1 dx
2 + 0 (1 + x)∆+
18 This behavior is expected at first glance since K vanishes if z → 0, but it diverges if also (x − x0 ) → 0.

16
substituing x = t2 ,
1
= C∆ z d−∆+ Ωd−1 B(d/2, ∆ − d/2)
2 +
1 2π d/2 Γ(d/2)Γ(∆ − d/2)
= C∆+ z d−∆+ .
2 Γ(d/2) Γ(∆)

So let’s choose the normalization constant C∆+ to be

Γ(∆)
C∆+ = . (5.18)
π d/2 Γ(ν)

K is not exactly a delta function as z → 0 since

Z
dd x K(z, x) = z ∆− , (5.19)

but z −∆− K is:

lim z −∆− K(z, x; x0 ) = δ (d) (x − x0 ). (5.20)
z→0

So near the boundary K has the behavior [19, lec. 14]

ε∆+
K(ε, x; x0 ) = ε∆− δ (d) (x − x0 ) + C∆+ . (5.21)
(x − x0 )2∆+

The first term should be regarded as letting possible to connect the bulk to the boundary.
Note that for a massless field, ∆− = 0 and K becomes really a delta function [33]. If we
Fourier transform the general expression (5.12) then we identify the same structure, and we
will see that coefficients agree exactly.
We can see that this propagator gives again the relation (2.29) between boundary and
bulk fields:
Z
φ(z, x) = C∆+ dd x0 K(z, x; x0 )φ0 (x0 ) =⇒ φ(ε, x) = ε∆− φ0 (x).

Now we have to derived the behavior of K under rescaling. Using the formula [11, sec.
23.5]
1 z 02 w02
= , (5.22)
(z − w)2 (z 0 − w0 )2
where w, z are adS coordinates and w0 , z 0 their transformations, we find

K(z, x1 ; x2 ) −→ (x01 )2∆+ K(z, x01 ; x02 ). (5.23)

6 Boundary-to-boundary propagator
It is possible to define a boundary-to-boundary propagator β(x; x0 ) which coincides with the
tree-level two-point function [28]. It is defined similarly as the bulk-to-boundary through a
double limit:
β(x; x0 ) = (2ν)2 lim
0
(zz 0 )−∆+ G(z, x; z 0 , x0 ). (6.1)
z,z →0

17
7 Full solution from Witten’s Green function
We have seen that near the boundary, the free field solution behaves as (2.27)
φfree (z, x) ∼0 z ∆− φ0 (x), (7.1)
and we will do as if we were starting from this fact. Since this is equivalent to have a source
on the boundary, we have to use the propagator to find a particular solution φ̄(z, x), using
here the Witten’s Green function: (5.17):
Z Z  2
z
φ̄(z, x) = d x K(z, x; x )φ0 (x ) = d x
d 0 0 0 d 0
φ0 (x0 ) (7.2)
z 2 + (x − x0 )2
which near the boundary behaves as
φ0 (x0 )
Z
φ̄(z, x) ∼0 C∆+ z ∆+ dd x0 . (7.3)
(x − x0 )2∆+
The complete solution of the Klein–Gordon equation with source is then
φ(z, x) = φfree (z, x) + φ̄(z, x) ∼0 z ∆− φ0 (x) + z ∆+ φ1 (x) (7.4)
with
φ0 (x0 )
Z
φ1 (x) = C∆+ dd x0 . (7.5)
(x − x0 )2∆+
Note that this last term, and then the complete solution, is not a local function of φ0 (x)
because of the integration. This asymptotic form is similar to the one discover computing
directly the solution (2.16).
Note that this relation can be rewritten
z ∆+
Z  
φ(z, x) ∼0 d x z δ (x − x ) + C∆+
d 0 ∆− (d) 0
φ0 (x0 ) (7.6)
(x − x0 )2∆+
and in the parenthesis we identify the form (5.21) of the propagator K near the boundary.

8 Witten’s diagrams
The computation of actions can be summarized by using Witten diagrams: as we have seen
the euclidean adS space is equivalent to a ball. With this last representation the boundary
of adS space maps to the boundary of the ball, and the bulk corresponds to its interior.
Lines between points give propagators (figure 1) [8, 11].

9 Evaluation of the action

Before evaluating the action using different methods, let’s pause a second to focus on a
potential problem: we have introduced everywhere a cut-off ε and at the end we want to
take the limit ε → 0. But this limit should be take in a consistent manner in all formula since
this limit and other expansions (e.g. of Bessel functions) do not commute [35]. Especially
we will see that one has to add counter-term to regulate divergences and if the limit is not
taken correctly then we get a wrong coefficient for the correlation function [12, 4, app. A],
[18, sec. 2.2], [19, lec. 17], [30, sec. 4] 19 .
Basically there are two ways to take the limit:
19 When using the adS/CFT conjecture one can relate 2- and 3-point functions, which determine the

"correct" coefficient [12, end of sec. 3.2], which is otherwise not possible in "normal" adS.

18
 (a) Empty adS. (b) Bulk-to-boundary (c) Bulk-to-bulk prop-
propagator. agator.

Figure 1: Witten’s diagram for empty adS and propagators.

• first expand all φ as φ ∼ φ0 + φ1 using (2.16), and then after taking the limit after
simplication (Witten’s and alternative momentum methods);
• or do simplifications and then take the limit ("Freedman’s" method).

9.1 Position space: Witten’s Green function

We will evaluate the scalar field action (1.1) using the second form (1.2): the first term
vanishes due to the equation of motion (1.4) and the second, as written in (1.3), reads

1 √
Z
Sε = dd y γ φnµ ∂µ φ
2 ∂M
Ld−1
Z 
= dd x z −d+1 φ(z, x)∂z φ(z, x)

2 z=ε

and using the expression (5.2) for then normal vector, and one introduces the boundary
propagator (2.27):

Ld−1
Z
S= dd x1 dd x2 φ0 (x1 )Fε (x1 , x2 )φ0 (x2 ) (9.1)
2
where Z 
Fε (x1 , x2 ) = z −d dd x K(z, x; x1 ) z∂z K(z, x; x2 ) . (9.2)

z=ε

Using the asymptotic expression (5.21) one can compute this operator to subleading or-
der [19, lec. 14]:

z ∆+
Z  
Fε (x1 , x2 ) ≈ z −d dd x z ∆− δ (d) (x − x1 ) + C∆+
(x − x1 )2∆+
z ∆+
 
× ∆− z ∆− δ (d) (x − x2 ) + ∆+ C∆+

 ,
(x − x2 )2∆ + 
z=ε

and after simplication (the fourth term vanishes in the limit ε → 0) 20

d C∆+
Fε (x1 , x2 ) ≈ ∆− ε−ν δ (d) (x1 − x2 ) + , (9.3)
(x1 − x2 )2∆+
20 Note that McGreevy [19] has a factor ∆
+ instead of d because he is taking the limit ε → 0 too soon,
approximating K(x1 ) by a delta function.

19
remembering that ∆+ + ∆− = d and d − 2∆− = 2ν. The first term is a contact term which
can be removed by renormalization, by adding a counter-term proportional to φ20 on the
boundary [20, sec. 4.3], [19, lec. 14], [34, p. 19], [27]. Then the correct expression is 21

Ld−1 ∆− √
Z
Sct = − dd x γ φ(z, x)2 , (9.4)
2 z=ε

because we have to use 5-dimensional fields to respect covariance [30, sec. 5.3] 22 .
Here the use of the Witten’s Green function is equivalent to first expanding field and
then taking the limit, so we have to do the same in Sct . Using the asymptotic expression
(2.16) (in position space) one gets

Ld−1 ∆−
Z  
Sct ≈ − dd x z −2ν φ0 (x)2 + 2φ0 (x)φ1 (x) . (9.5)
2 z=ε

Using the expression (7.5) of φ1 in term of K we rewrite the second term as

Ld−1 ∆− C ∆+
Z
− dd x1 dd x2 2φ0 (x1 )φ0 (x2 ) ,
2 z=ε (x1 − x2 )2∆+

which results in the total action [34, p. 19]

Ld−1
Z
S= dd x1 dd x2 φ0 (x1 )Feε (x1 , x2 )φ0 (x2 ) (9.6)
2
where

e 1 , x2 ) = Fε (x1 , x2 ) − ∆− z −2ν δ (d) (x1 − x2 ) − 2∆− C ∆+

F(x (9.7a)
(x1 − x2 )2∆+
2ν C∆+
= (9.7b)
(x1 − x2 )2∆+

is the renormalized kernel: it is independent of ε.

The final renormalized action is then [26, sec. 4.2]

φ0 (x1 )φ0 (x2 )

Z
S[φ0 ] = ν Ld−1 C∆+ dd x1 dd x2 . (9.8)
(x1 − x2 )2∆+

In section 3 we said that φ1 is the conjugate variable to φ0 . We are now able to prove
this assertion by computing the derivative of S[φ0 ] [22]

δS[φ0 ] φ0 (x0 )
Z
= ν L C∆+ dd x0
d−1
= ν Ld−1 φ1 (x) (9.9)
δφ0 (x) (x − x0 )2∆+

using the relation (7.5).

9.2 Momentum space: Green function ("Freedman’s" method)

The previous method is not always available, for example when the interior is not adS
anymore. So we have to work in momentum space, using the expression (2.34) (as we note
earlier it is almost the propagator, so we will save some pain by using this expression instead
21 Some people are using directly the boundary field, giving a different factor.
22 Note also that this boundary term ensures stability[27].

20
of the propagator). The insertion of the Fourier transform (2.7) in the action (1.3) gives [11,
sec. 23.10]
Ld−1
Z 
S[φ0 ] = dd x z −d+1 φ(z, x)∂z φ(z, x)

2 z=ε

Ld−1 Z
d k d k
d d 0 0

= dd x eix(k+k ) z −d+1 fk (z)∂z fk0 (z)

2 (2π) (2π)
d d z=ε

Ld−1 Z
d k d k
d d 0
ix(k+k0 )

= dd
x e φ (k)φ (k 0
) z −d+1
K (z)∂ K 0 (z)

0 0 k z k
2 (2π)d (2π)d z=ε

and using (5.3)

Ld−1 dd k
Z
S= φ0 (k) Fε (k) φ0 (−k) (9.10)
2 (2π)d
where 
Fε (k) = z −d K−k (z) z∂z Kk (z) . (9.11)

z=ε
By plugging (5.11) this can be rewritten 23

d  
Fε (k) = ε2∆− ε−d+1 ln (kε)d/2 Kν (kε) . (9.12)
dε
In the parenthesis we recognize fk (ε)/ak from (2.15) so we can use the asymptotic (2.16) to
compute the derivative 24 :
d d 1
    
ε ln ε Kν (kε) ≈ ε ln
d/2
φ0 ε ∆−
+ φ1 ε ∆+
dε dε ak
d
  
φ1 2ν
=ε − ln ak + ln φ0 + ∆− ln ε + ln 1 + ε
dε φ0
so expand the logarithm and take to zero the derivative of the two constants,
φ1 2ν
= ∆− + 2ν ε .
φ0
The first term is analytic in k (if we were going further in the development, there would be
only integer power of k), so it corresponds to contact term and it is removed by renormal-
ization: the corresponding counter-term is the one written previously (9.4). Note that we
have to take the limit ε → 0 directly, using the limit (2.27):
Ld−1 ∆− √
Z
Sct ≈ − dd x γ ε2∆− φ0 (x)2 . (9.13)
2 z=ε

The second term is non-analytic and it corresponds to the absorptive part of the two-
point function:
 2ν
Γ(−ν) k
Fε (k) = 2ν (9.14)
Γ(ν) 2
where we used (2.20), and the ε2∆− factor cancels the ε−2∆− appearing in the product 25 .
So finally the action is
 2ν
Γ(−ν) dd k
Z
k
S[φ0 ] = ν Ld−1 φ 0 (−k)φ 0 (k) . (9.15)
Γ(ν) (2π) d 2
23 We add a factor (k/k)d/2 for convenience.
24 Freedman [11, p. 548] directly expands the Bessel function, the result is the same but our way avoid
expanding again.
25 As we said in section 2.5 most of people use the boundary condition at z = 0 instead of z = ε, leading

to a factor ε−2∆− , which they throw away saying that one has to rescale the field.

21
Transforming this back to position space gives

Γ(−ν) dd k 2ν ik(x1 −x2 )

Z Z
S[φ0 ] = ν Ld−1 2ν dd x1 dd x2 φ0 (x1 )φ0 (x2 ) k e
2 Γ(ν) (2π)d
Γ(−ν) 22ν Γ(∆+ ) 1
Z
= ν Ld−1 2ν dd x1 dd x2 φ0 (x1 )φ0 (x2 ) d/2 ,
2 Γ(ν) π Γ(−ν) (x1 − x2 )2∆+
and finally
φ0 (x1 )φ0 (x2 )
Z
S[φ0 ] = ν Ld−1 C∆+ dd x1 dd x2 . (9.16)
(x1 − x2 )2∆+
We get the same result as in the previous section.
The computation for ν integer will be multiplied by a logarithm.

9.3 Momentum space: Green function (alternative method)

If we compute the kernel Fε (k) (9.11) without writing it as a log and use (5.12), then we
need to use the action (9.5) where φ is expanded [34, p. 19]:

Fε (k) = z −d K−k (z) z∂z Kk (z)

  z=ε  
φ 1 ∆+ φ1 ∆+ 
=z −d
z ∆−
+ z z∂z z ∆− + z
φ0 φ0

z=ε
φ1
= ∆− ε2ν + d .
φ0
Here the factor of the second term is d, so subtracting 2∆− gives the expected 2ν.

10 Cubic interactions
To the free action (1.1) for three fields φi
1 √
Z
S0 = − dd+1 X g (g µν ∂µ φi ∂ν φi + m2ij φi φj ), mij = mi δij , (10.1)
2
we add a cubic interaction
√
Z
Sint = − dd+1 X g λijk φi φj φk . (10.2)

This gives the equation of motion

(− ∆ +m2i )φi = λijk φj φk . (10.3)

The solution for φi is obtained perturbatively by introducing more and more source on the
boundary and interactions in the bulk:
Z
φi (z, x) = dd x1 Ki (z, x; x1 )φi0 (x1 )
X Z
+ λijk dd+1 X 0 Gi (z, x; z 0 , x0 )×
j,k (10.4)
Z
× dd x1 dd x2 Kj (z 0 , x0 ; x1 )Kk (z 0 , x0 ; x2 )φ0j (x1 )φ0k (x2 )

+ ···

22
Each term in this expansion can be represented as a Witten diagram.
Now we can compute the action, keeping only terms with three sources; the free action
part will not contribute and so only the first term in the φi expansion is needed:
Z
(3)
Sint = dd x1 dd x2 dd x3 Fijk (x1 , x2 , x3 )φ0i (x1 )φ0j (x2 )φ0k (x3 ) (10.5)

where Z
Fijk (x1 , x2 , x3 ) = λijk dd+1 X Ki (z, x; x1 )Kj (z, x; x2 )Kk (z, x; x3 ) (10.6)

(no summation). Note that the interaction part can not give any contribution to the term
quadratic in the sources, so the result we found in the free theory for this quadratic part
is still valid. Moreover there are no ε left so we do not need renormalization [23]. The
corresponding Witten’s diagram is given in figure 2 (bulk-to-bulk propagator appears with
four sources).

Figure 2: Witten’s diagram for cubic interaction.

Question 10.1 Check that the free part does not contribute. Add references. Cf
Kiritsis [15, p. 429, 551].

11 Comments
Generalization to the Lifschitz space corresponding to the metric
dt2 dz 2 1
ds2 = − + + 2 δij dxi dxj , (11.1)
z 2z z2 z
where the symmetries on the boundary lie in the Schrödinger group (non-relativistic scale
invariant theory) can be found in [2, 31, 13]; z is the dynamical exponent which gives the
anisotropy between time and space under scale transformation
t −→ λz , xi −→ λxi . (11.2)
z = 1 corresponds to full conformal symmetry and z = 2 to the Schrödinger equation.
Computations of n-point functions with arbitrary polynomial can be found in [23].
In all computations we used Poincaré coordinates as a vacuum. Global coordinates are
also often used, and sometimes more convenient. We know that the vacuum choice can have
an impact when quantizing a field theory on curved space; the authors of [3] relate modes
of different vacua (using adS/CFT) and show that the only difference is an energy shift of
the states.

23
A Conventions
A.1 Basic
The metric signature is taken to be mostly plus.
We define by g the absolute value of the metric determinant:

g = |det gµν | = − det gµν . (A.1)

A.2 Wick rotation

The passage from Minkowski metric

ds2 = ηµν dxµ dxν = −dt2 + dx2 (A.2)

to euclidean metric
ds2E = δµν dxµE dxνE = dτ 2 + dx2 (A.3)
is done through the substitution of the real time t by the euclidean time τ [36, sec. 3.4]

t = −iτ. (A.4)

Then the action becomes [32]

Z Z Z
S = dd x L = −i dd xE L = i dd xE LE = iSE (A.5)

so that Z Z
S = iSE , L = −LE , Z= dφ eiS = dφ e−SE (A.6)

Since the Euclidean action is now positive definite, the minus sign in the partition function
gives exponential damping. In the case of curved spacetime the metric determinant simply
becomes √ √
−g = gE . (A.7)
As an example look at the scalar lagrangian with potential:
1 
L = − (∂ µ φ)2 + m2 φ2 − V (φ) (A.8)
2
which gives the equation of motion

(− ∆ +m2 )φ = V 0 (φ). (A.9)

Plugging plane-waves into the free equation (V = 0) gives the mass-shell condition

p2 = −m2 (A.10)

and the Green function

1
G(p) = (A.11)
p2 + m2
has a singularity.
Applying the Wick rotation gives
1 µ 2 
LE = (∂E φ) + m2 φ2 + V (φ), (A.12)
2

24
which is positive definite, and the equation of motion

(− ∆E +m2 )φ = −V 0 (φ). (A.13)

with Green function

1
G(p) = . (A.14)
p2E + m2
This function has no singularity since plane-waves are not anymore solutions of the Klein-
Gordon equation (said another way, there is no particle in Euclidean space).

B Anti de Sitter space

This apppendix mainly draw from [11, sec. 22.1], and [5] is another useful review.
In this chapter we will use as indices:
• A for (d + 1)-dimensional (adS);
• µ for d-dimensional Minkowski;
• i for (d − 1)-dimensional spatial.

In some cases we will also consider the embedding into (d+2)-dimensional space with indices
α, and we will note a = {i}, d.

B.1 Action and equation of motion

Start with the Einstein–Hilbert action in vacuum with cosmological constant Λ in D = d + 1
dimensions
1 √
Z
S= 2 dd+1 x −g(R − Λ). (B.1)
2κ
If we write
1
Λ=− d(d − 1) (B.2)
L2
the equation of motion reads
d 1
Rµν = − gµν , R=− d(d + 1), (B.3a)
L2 L2
which means that the space is an Einstein manifold, and moreover it is maximally symmet-
ric [7, p. 141]:
1
Rµνρσ = − 2 (gµρ gνσ − gµσ gνρ ). (B.3b)
L
L will be the radius of the space (equivalently size of the throat).

B.2 Coordinates
B.2.1 Global
Let’s denote generically the metric by

ds2 = gAB dX A dX B (B.4)

where the A indices are splitten between a radial coordinate and d-dimensional µ = 0, . . . , d−
1 (in some cases they will correspond to Minkowski indices).

25
 From the embedding into Rd,2 with

ηαβ = diag(−1, 1, . . . , 1, −1), α = 0, . . . , d + 1, (B.5)

adSd+1 is defined as an hyperboloid of radius L

ηαβ Y α Y β = ηµν Y µ Y ν + (Y d )2 − (Y d+1 )2 = −L2

= −(Y 0 )2 + (Y a )2 − (Y d+1 )2 = −L2 ,

where we have denoted a = {i}, d and as usual i = 1, . . . , d − 1.

The choice of parametrization
p t p t
Y0 = L2 + r2 sin , Y d+1 = L2 + r2 cos , Y a = rx̂a (B.7)
L L
with
r ≥ 0, t ∈ [0, 2πL[, (x̂a )2 = 1 (B.8)
gives the metric
−1
r2 r2
  
ds = − 1 + 2
2
dt + 1 + 2
2
dr2 + r2 dΩ2d−1 . (B.9)
L L

The time is periodic and closed time-like curved exists. To get rid of this unwanted feature
one can take the covering space where t ∈ R – we do not use a new name but implictly we
use this covering space. In these coordinates the SO(d, 2) symmetry is manifest [27].
The change of radial variable
r
y r y r2
sh = , ch = 1 + 2 , y≥0 (B.10)
L L L L
gives
y 2 y
ds2 = − ch2dt + dy 2 + L2 sh2 dΩ2d−1 . (B.11)
L L
From this last coordinate system we can introduce again a new radial coordinates and a
rescaled time
y 1
t = Lτ, ch = , ρ ∈ [0, π/2[, (B.12)
L cos ρ
which gives the metric

L2
ds2 = −dτ 2 + dρ2 + sin2 ρ dΩ2d−1 . (B.13)


cos ρ
2

This metric is a conformal factor times another metric known as the Einstein static universe,
which is R × S d−1 ; but since to cover the whole sphere one needs ρ ∈ [0, π[, only the upper
half plane is covered here.
The Cauchy problem is ill-defined in adS space because information from spatial infinity
can reach the origin in finite time: it will explain why in adS/CFT one needs to give
boundary conditions.

26
B.2.2 Patches
Another set of very useful coordinates is the Poincaré patch, given by

L2 2 r2
ds2 = dr + 2 ηµν dxµ dxν , r > 0, (B.14)
r2 L
or by rescaling r = Lu:
du2
 
ds2 = L2 + u 2
η µν dx µ
dxν
. (B.15)
u2
Note that only a part of the adS space is covered by these coordinates, and the r → 0
limit is an horizon, not a singularity. For the euclidean version r = 0 shrinks to a point.
In the case og minkowskian signature we will need to consider ingoing or outgoing waves as
boundary conditions when solving the wave equations.
The change of variables
L2 1
z= = (B.16)
r u
brings the metric to the form

L2  2 
ds2 = dz + η µν dx µ
dxν
. (B.17)
z2
The boundary is at z → 0.
Null geodesics are given by [27]

z = ±(t − t0 ) (B.18)

for which the boundary z = 0 is reached in a finite time, while it takes an infinite time to
reach z = ∞.
The normal vector to a surface z = cst is
1 z
n= √ ∂z = ∂z (B.19)
gzz L

(normalized such that n2 = 1) and the induced metric is

L2
γµν = ηµν . (B.20)
z2

B.3 Properties
Given two points X = (z, x) and X 0 = (x0 , z 0 ), the unique conformal invariant that can be
constructed is [27]
2zz 0
ξ= 2 . (B.21)
z + z 02 + (x − x0 )2

B.4 Boundary
Consider the equation of embedding (B.6)

− (Y 0 )2 + (Y a )2 − (Y d+1 )2 = −L2 (B.22)

and introduce light-cone coordinates [1, sec. 4.1.1]

U = Y d+1 + Y d , V = Y d+1 − Y d . (B.23)

27
The constraint equation becomes

− U V − (Y 0 )2 + (Y i )2 = −L2 , i = 1, . . . , d − 1. (B.24)

If L = 0 then one can solve for V if U 6= 0 as

−(Y 0 )2 + (Y a )2
V = . (B.25)
U
Note that we also have the symmetry where Y α → λ Y α , and we mod it out by using it
to fix one of the coordinates, say U = 1 (this is equivalent to identifying two points if they
differ by a rescaling). The induced metric is then d-dimensional Minkowski

ds2 = ηµν dxµ dxν . (B.26)

However we can also include the point U = 0 by reversing the roles of U and V ; concretly
we have added the point at infinity and the space we get has not exactly the topology of
Minkowski but the one of (S 1 × S d−1 )/Z2 , or R × S d−1 if we consider the universal covering.
It is not possible to define directly the induced metric on the boundary z = 0 because it
is divergent, so we define it as the limit [1, sec. 5.2] (omitting the L2 factor)

f2
γµν = lim ηµν (B.27)
z→0 z 2

for some function f such that f (z) ∼0 z. Thus the induced metric depends on f and is
defined up to a conformal transformation, which means that we have a conformal structure
on the boundary (i.e. we can measure angles but not distances) 26 . Doing a conformal
transformation f → ew f implies
0
γµν = e2w γµν , (B.28)
so the metric has conformal weight −2.

C AdS/CFT correspondence
C.1 The dictionnary
The basic correspondence says that to each boundary field there is an associated boundary
operator from a CFT. The generating function is given by
 Z 
ZCFT = exp φ0 O (C.1)
∂M

and one postulates that it is given by the extremum of the string theory action, i.e. at tree
level it corresponds to the action of supergravity evaluated with classical configuration:

ZCFT = exp(−S[φ0 ]) (C.2)

where the φ satisfy the asymptotic condition (2.27).

If φ couples to an operator O on the boundary, we can use the following trick to find the
scaling of this operator [19]:
 d
√
Z Z
L
S3 dd x γε φ(ε, x)O(ε, x) = dd x ε∆− φ0 (x)O(x)
z=ε z=ε ε
26 This is expected in context of adS/CFT, where the CFT lives on the boundary.

28
where we used the boundary field (2.29)

φ(ε, x) = ε∆− φ0 (x), (C.3)

defined in the previous section, and so

O(ε, x) = ε∆+ O(x). (C.4)

C.2 Choice of coordinate systems

A question that can be asked concerns the choice between the global coordinate system
(B.9)
−1
r2 r2
  
ds2 = − 1 + 2 dt2 + 1 + 2 dr2 + r2 dΩ2d−1 (C.5a)
L L
and the Poincaré upper-half plane (B.17)

L2  2 
ds2 = dz + ηµν dx µ
dx ν
. (C.5b)
z2
The later presents an horizon at z = ∞, but we will study only systems with finite time
and length, so the horizon will never be crossed.
It is worth to note that the dual CFT is defined on S d−1 instead of Rd−1 [27].

C.3 Correlation functions

C.3.1 1-point
C.3.2 2-point
Using (9.16) the 2-point function is given by

1

δ δ
hO(x1 )O(x2 )i = − S[φ0 ] = −ν Ld−1 C∆+ (C.6)

δφ0 (x1 ) δφ0 (x2 ) φ0 =0 (x1 − x2 )2∆+

which agrees with the result one would expect from conformal invariance.

C.3.3 3-point
If we turn on a cubic interaction in the bulk for three fields φi as in (10.2), then the 3-
point function for the associated operators Oi is given by deriving (10.5) with respect to the
sources:
hOi (x1 )Oj (x2 )Ok (x3 )i = Fijk (x1 , x2 , x3 ) (C.7)
with Z
Fijk (x1 , x2 , x3 ) = λijk dd+1 x Ki (z, x; x1 )Kj (z, x; x2 )Kk (z, x; x3 ) (C.8)

Using the Witten’s expression for Ki and the scaling relation (5.23)

K(z, x1 ; x2 ) −→ (x01 )2∆+ K(z, x01 ; x02 ) (C.9)

one finds that

hOi (x1 )Oj (x2 )Ok (x3 )i −→ (x01 )2∆i (x02 )2∆j (x03 )2∆k hOi (x01 )Oj (x02 )Ok (x03 )i (C.10)

which is the correct transformation for a conformal 3-point function.

29
D Special functions
D.1 Gamma and beta functions
The beta function is Z 1
B(x, y) = dt tx−1 (1 − t)y−1 (D.1)
0
We can obtain the equivalent forms
∞
tx−1 Γ(x)Γ(y)
Z
B(x, y) = dt = (D.2)
0 (1 + t)x+y Γ(x + y)

D.2 Bessel functions

Bessel differential equation is

x2 f 00 + xf 0 − (x2 − n2 )f = 0 (D.3)

admits J±n (Bessel functions of first kind) as solutions if n ∈

/ N.
If n ∈ N then J±n (x) are not linearly independent since

J−n = (−1)n Jn . (D.4)

In this case one has to introduce Bessel functions of second kind Yn (x) (denoted sometime
Nn (x))
Jn (x) cos(nx) − J−n (x)
Yn (x) = , (D.5)
sin(nx)
the limit where n is an integer being regular.
Asymptotic forms are pure imaginary exponentials.

D.3 Modified Bessel functions

Modified Bessel functions In (x) and Kn (x) are solutions of the differential equation

x2 f 00 + xf 0 − (x2 + n2 )f = 0. (D.6)

The modified Bessel functions have exponential behavior.

One has the asymptotic forms
 n
1  x n Γ(n) 2
In (x) ∼0 , Kn (x) ∼0 , (D.7a)
Γ(n + 1) 2 2 x
ex e−x
In (x) ∼∞ √ , Kn (x) ∼∞ √ ? (D.7b)
x x
The derivative of the Bessel function is
dKn (x) 1 
= − Kn−1 (x) + Kn+1 (x) , (D.8)
dx 2
and
dKn (x)
x = nKn (x) − x Kn+1 (x). (D.9)
dx
The complete limits for Kn is [8, p. 548]
 n
Γ(n) 2 Γ(−n)  x n
Kn (x) ∼0 + . (D.10)
2 x 2 2

30
If n is integer, then there is an extra factor ln x in the second term.
In terms of Bessel functions they are given as

In (x) = i−n Jn (ix),

π I−n (x) − In (x) (D.11)
Kn (x) = .
2 sin(nx)

31
References
[1] Steven G. Avery. adS/CFT correspondence. 2008.
url: http://www.imsc.res.in/~savery/papers/paper.pdf.
[2] Koushik Balasubramanian and John McGreevy. Gravity duals for non-relativistic CFTs.
arXiv e-print 0804.4053. Apr. 2008.
[3] Vijay Balasubramanian, Per Kraus, and Albion Lawrence. “Bulk vs. Boundary Dy-
namics in Anti-de Sitter Spacetime” (May 1998).
doi: 10.1103/PhysRevD.59.046003.
arXiv: hep-th/9805171.
[4] Vijay Balasubramanian et al. “Holographic Probes of Anti-de Sitter Spacetimes” (Aug.
1998).
doi: 10.1103/PhysRevD.59.104021.
arXiv: hep-th/9808017.
[5] Ingemar Bengtsson. Anti de Sitter space. 1998.
url: http://www.physto.se/~ingemar/Kurs.pdf.
[6] Peter Breitenlohner and Daniel Z Freedman. “Stability in gauged extended supergrav-
ity”. Annals of Physics 144.2 (Dec. 1982).
doi: 10.1016/0003-4916(82)90116-6.
[7] Sean M Carroll. Spacetime and geometry: an introduction to general relativity. English.
Addison Wesley, 2004.
[8] Eric D’Hoker and Daniel Z Freedman. “Supersymmetric Gauge Theories and the
AdS/CFT Correspondence” (Jan. 2002).
arXiv: hep-th/0201253.
[9] Hans Günter Dosch. A Practical Guide to AdS/CFT. 2009.
url: http://www.thphys.uni-heidelberg.de/~dosch/wuhantot.pdf.
[10] Michael R. Douglas and S. Randjbar-Daemi. “Two Lectures on the AdS/CFT Corre-
spondence” (Feb. 1999).
arXiv: hep-th/9902022.
[11] Daniel Z Freedman and Antoine Van Proeyen. Supergravity. English. Cambridge Uni-
versity Press, 2012.
[12] Daniel Z. Freedman et al. “Correlation functions in the CFT(d)/AdS(d+1) correpon-
dence” (Apr. 1998).
doi: 10.1016/S0550-3213(99)00053-X.
arXiv: hep-th/9804058.
[13] Tom Griffin, Petr Horava, and Charles M. Melby-Thompson. “Lifshitz Gravity for
Lifshitz Holography” (Nov. 2012).
arXiv: 1211.4872.
[14] S. S. Gubser, I. R. Klebanov, and A. M. Polyakov. “Gauge Theory Correlators from
Non-Critical String Theory” (Feb. 1998).
doi: 10.1016/S0370-2693(98)00377-3.
arXiv: hep-th/9802109.
[15] Elias Kiritsis. String Theory in a Nutshell. English. Princeton University Press, 2007.
[16] Ingo Kirsch. Introduction to AdS/CFT. 2012.
url: http://marcobaumgartl.info/sfp12/scripts/AdSCFT.pdf.

32
[17] Igor R. Klebanov. “TASI Lectures: Introduction to the AdS/CFT Correspondence”
(Sept. 2000).
arXiv: hep-th/0009139.
[18] Igor R. Klebanov and Edward Witten. “AdS/CFT Correspondence and Symmetry
Breaking” (May 1999).
doi: 10.1016/S0550-3213(99)00387-9.
arXiv: hep-th/9905104.
[19] John McGreevy. Applied String Theory. 2008.
url: http : / / ocw . mit . edu / courses / physics / 8 - 821 - string - theory - fall -
2008/lecture-notes/.
[20] John McGreevy. “Holographic duality with a view toward many-body physics” (Sept.
2009).
arXiv: 0909.0518.
[21] Pablo Minces and Victor O. Rivelles. “Scalar Field Theory in the AdS/CFT Corre-
spondence Revisited” (July 1999).
doi: 10.1016/S0550-3213(99)00833-0.
arXiv: hep-th/9907079.
[22] W. Mück and K. S. Viswanathan. Regular and Irregular Boundary Conditions in the
AdS/CFT Correspondence. arXiv e-print hep-th/9906155. June 1999.
arXiv: hep-th/9906155.
[23] Wolfgang Mück. “Studies on the adS/CFT correspondence”. PhD thesis. Simon Fraser
University, 1999.
url: http://www.collectionscanada.gc.ca/obj/s4/f2/dsk1/tape9/PQDD_0025/
NQ51904.pdf.
[24] Wolfgang Mück and K. S. Viswanathan. “Conformal Field Theory Correlators from
Classical Scalar Field Theory on AdS_{d + 1}” (Apr. 1998).
doi: 10.1103/PhysRevD.58.041901.
arXiv: hep-th/9804035.
[25] Juan F. Pedraza and Pavel Safronov. The adS/CFT correspondence. 2012.
url: http://www.ma.utexas.edu/users/psafronov/notes/adscft.pdf.
[26] Jens L. Petersen. “Introduction to the Maldacena Conjecture on AdS/CFT” (Feb.
1999).
doi: 10.1142/S0217751X99001676.
arXiv: hep-th/9902131.
[27] Joseph Polchinski. “Introduction to Gauge/Gravity Duality” (Oct. 2010).
arXiv: 1010.6134.
[28] Karl-Henning Rehren. “QFT Lectures on AdS-CFT” (Nov. 2004).
arXiv: hep-th/0411086.
[29] Yuji Satoh and Jan Troost. “On time-dependent AdS/CFT”. Journal of High Energy
Physics 2003.01 (Jan. 2003), pp. 027–027.
doi: 10.1088/1126-6708/2003/01/027.
[30] Kostas Skenderis. “Lecture Notes on Holographic Renormalization” (Sept. 2002).
doi: 10.1088/0264-9381/19/22/306.
arXiv: hep-th/0209067.
[31] D. T. Son. “Toward an AdS/cold atoms correspondence: a geometric realization of the
Schroedinger symmetry” (Apr. 2008).
doi: 10.1103/PhysRevD.78.046003.
arXiv: 0804.3972.

33
[32] D. T Son and A. O Starinets. “Minkowski-space correlators in AdS/CFT correspon-
dence: recipe and applications” (May 2002).
doi: 10.1088/1126-6708/2002/09/042.
arXiv: hep-th/0205051.
[33] Edward Witten. “Anti De Sitter Space And Holography” (Feb. 1998).
arXiv: hep-th/9802150.
[34] Jan Zaanen et al. The AdS/CMT manual for plumbers and electricians. 2012.
url: http://www.solvayinstitutes.be/events/doctoral/J.Zaanen_Notes.pdf.
[35] Alberto Zaffaroni. Introduction to the AdS/CFT correspondence. 2009.
url: http://laces.web.cern.ch/Laces/LACES09/notes/dbranes/lezioniLosanna.
pdf.
[36] Alexei Zamolodchikov and Alexander Zamolodchikov. Lectures on Liouville Theory
and Matrix Models. 2007.
url: http://qft.itp.ac.ru/ZZ.pdf.

34